Skip to contentSkip to conversation

Fix DPC hotplug race and enhance error handling

PATCH Superseded Ethan Zhao 5 participants 12 messages Oct 3, 2020 – Oct 7, 2020

Subsystems (2)

v7 of 8 Superseded Latest is v8 →

lore
Share & tools
lore

Ethan Zhao

author
#0 lore raw

v7 cover Fix DPC hotplug race and enhance error handling

Hi,folks,

This simple patch set fixed some serious security issues found when DPC error injection and NVMe SSD hotplug brute force test were doing -- race condition between DPC handler and pciehp, AER interrupt handlers, caused system hang and system with DPC feature couldn't recover to normal working state as expected (NVMe instance lost, mount operation hang, race PCIe access caused uncorrectable errors reported alternatively etc).

With this patch set applied, stable 5.9-rc6 on ICS (Ice Lake SP platform, see
https://en.wikichip.org/wiki/intel/microarchitectures/ice_lake_(server))

could pass the PCIe Gen4 NVMe SSD brute force hotplug test with any time interval between hot-remove and plug-in operation tens of times without any errors occur and system works normal.

With this patch set applied, system with DPC feature could recover from NON-FATAL and FATAL errors injection test and works as expected.

System works smoothly when errors happen while hotplug is doing, no uncorrectable errors found.

Brute DPC error injection script:

for i in {0..100}
do

        setpci -s 64:02.0 0x196.w=000a
        setpci -s 65:00.0 0x04.w=0544
        mount /dev/nvme0n1p1 /root/nvme
        sleep 1

done

Other details see every commits description part.

This patch set could be applied to stable 5.9-rc6/rc7 directly.

Help to review and test.

v2: changed according to review by Andy Shevchenko.
v3: changed patch 4/5 to simpler coding.
v4: move function pci_wait_port_outdpc() to DPC driver and its
declaration to pci.h. (tip from Christoph Hellwig <[email protected]>).
v5: fix building issue reported by [email protected] with some config.
v6: move patch[3/5] as the first patch according to Lukas's suggestion.
and rewrite the comment part of patch[3/5].
v7: change the patch[4/5], based on Bjorn's code and truth table.
change the patch[5/5] about the debug output information.

Thanks,
Ethan

Ethan Zhao (5):
PCI/ERR: get device before call device driver to avoid NULL pointer
dereference
PCI/DPC: define a function to check and wait till port finish DPC
handling
PCI: pciehp: check and wait port status out of DPC before handling
DLLSC and PDC
PCI: only return true when dev io state is really changed PCI/ERR: don't mix io state not changed and no driver together

 drivers/pci/hotplug/pciehp_hpc.c |  4 ++-
 drivers/pci/pci.h                | 55 +++++++++++++-------------------
 drivers/pci/pcie/dpc.c           | 27 ++++++++++++++++
 drivers/pci/pcie/err.c           | 18 +++++++++--
 4 files changed, 68 insertions(+), 36 deletions(-)

base-commit: a1b8638ba1320e6684aa98233c15255eb803fac7

Signature · 1 line
2.18.4
author
#1 lore raw

PATCH v7 1/5 PCI/ERR: get device before call device driver to avoid NULL pointer dereference

During DPC error injection test we found there is race condition between pciehp and DPC driver, NULL pointer dereference caused panic as following

 # setpci -s 64:02.0 0x196.w=000a
  // 64:02.0 is rootport has DPC capability
 # setpci -s 65:00.0 0x04.w=0544
  // 65:00.0 is NVMe SSD populated in above port
 # mount /dev/nvme0n1p1 nvme

(tested on stable 5.8 & ICS(Ice Lake SP platform, see
https://en.wikichip.org/wiki/intel/microarchitectures/ice_lake_(server))

 BUG: kernel NULL pointer dereference, address: 0000000000000050
 ...
 CPU: 12 PID: 513 Comm: irq/124-pcie-dp Not tainted 5.8.0-0.0.7.el8.x86_64+ #1
 RIP: 0010:report_error_detected.cold.4+0x7d/0xe6
 Code: b6 d0 e8 e8 fe 11 00 e8 16 c5 fb ff be 06 00 00 00 48 89 df e8 d3 65 ff
 ff b8 06 00 00 00 e9 75 fc ff ff 48 8b 43 68 45 31 c9 <48> 8b 50 50 48 83 3a 00
 41 0f 94 c1 45 31 c0 48 85 d2 41 0f 94 c0
 RSP: 0018:ff8e06cf8762fda8 EFLAGS: 00010246
 RAX: 0000000000000000 RBX: ff4e3eaacf42a000 RCX: ff4e3eb31f223c01
 RDX: ff4e3eaacf42a140 RSI: ff4e3eb31f223c00 RDI: ff4e3eaacf42a138
 RBP: ff8e06cf8762fdd0 R08: 00000000000000bf R09: 0000000000000000
 R10: 000000eb8ebeab53 R11: ffffffff93453258 R12: 0000000000000002
 R13: ff4e3eaacf42a130 R14: ff8e06cf8762fe2c R15: ff4e3eab44733828
 FS:  0000000000000000(0000) GS:ff4e3eab1fd00000(0000) knlGS:0000000000000000
 CS:  0010 DS: 0000 ES: 0000 CR0: 0000000080050033
 CR2: 0000000000000050 CR3: 0000000f8f80a004 CR4: 0000000000761ee0
 DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000
 DR3: 0000000000000000 DR6: 00000000fffe0ff0 DR7: 0000000000000400
 PKRU: 55555554
 Call Trace:
 ? report_normal_detected+0x20/0x20
 report_frozen_detected+0x16/0x20
 pci_walk_bus+0x75/0x90
 ? dpc_irq+0x90/0x90
 pcie_do_recovery+0x157/0x201
 ? irq_finalize_oneshot.part.47+0xe0/0xe0
 dpc_handler+0x29/0x40
 irq_thread_fn+0x24/0x60
 ...

Debug shows when port DPC feature was enabled and triggered by errors, DLLSC/PDC/DPC interrupts will be sent to pciehp and DPC driver almost at the same time, and no delay between them is required by specification. so DPC driver and pciehp drivers may handle these interrupts cocurrently.

While DPC driver is doing pci_walk_bus() and calling device driver's callback without pci_dev_get() to increase device reference count, the device and its driver instance are likely being freed by pci_stop_and_removed_bus_device() -> pci_dev_put().

So does pci_dev_get() before using the device instance to avoid NULL pointer dereference.

---
Changes:
v2: revise doc according to Andy's suggestion.
v3: no change.
v4: no change.
v5: no change.
v6: moved to [1/5] from [3/5] and revised comment according to Lukas'
suggestion.
v7: no change.

 drivers/pci/pcie/err.c | 12 ++++++++++++
 1 file changed, 12 insertions(+)
drivers/pci/pcie/err.c +12 -0
index c543f419d8f9..e35c4480c86b 100644
@@ -52,6 +52,8 @@ static int report_error_detected(struct pci_dev *dev,
 	pci_ers_result_t vote;
 	const struct pci_error_handlers *err_handler;
 
+	if (!pci_dev_get(dev))
+		return 0;
 	device_lock(&dev->dev);
 	if (!pci_dev_set_io_state(dev, state) ||
 		!dev->driver ||
@@ -76,6 +78,7 @@ static int report_error_detected(struct pci_dev *dev,
 	pci_uevent_ers(dev, vote);
 	*result = merge_result(*result, vote);
 	device_unlock(&dev->dev);
+	pci_dev_put(dev);
 	return 0;
 }
 
@@ -94,6 +97,8 @@ static int report_mmio_enabled(struct pci_dev *dev, void *data)
 	pci_ers_result_t vote, *result = data;
 	const struct pci_error_handlers *err_handler;
 
+	if (!pci_dev_get(dev))
+		return 0;
 	device_lock(&dev->dev);
 	if (!dev->driver ||
 		!dev->driver->err_handler ||
@@ -105,6 +110,7 @@ static int report_mmio_enabled(struct pci_dev *dev, void *data)
 	*result = merge_result(*result, vote);
 out:
 	device_unlock(&dev->dev);
+	pci_dev_put(dev);
 	return 0;
 }
 
@@ -113,6 +119,8 @@ static int report_slot_reset(struct pci_dev *dev, void *data)
 	pci_ers_result_t vote, *result = data;
 	const struct pci_error_handlers *err_handler;
 
+	if (!pci_dev_get(dev))
+		return 0;
 	device_lock(&dev->dev);
 	if (!dev->driver ||
 		!dev->driver->err_handler ||
@@ -124,6 +132,7 @@ static int report_slot_reset(struct pci_dev *dev, void *data)
 	*result = merge_result(*result, vote);
 out:
 	device_unlock(&dev->dev);
+	pci_dev_put(dev);
 	return 0;
 }
 
@@ -131,6 +140,8 @@ static int report_resume(struct pci_dev *dev, void *data)
 {
 	const struct pci_error_handlers *err_handler;
 
+	if (!pci_dev_get(dev))
+		return 0;
 	device_lock(&dev->dev);
 	if (!pci_dev_set_io_state(dev, pci_channel_io_normal) ||
 		!dev->driver ||
@@ -143,6 +154,7 @@ static int report_resume(struct pci_dev *dev, void *data)
 out:
 	pci_uevent_ers(dev, PCI_ERS_RESULT_RECOVERED);
 	device_unlock(&dev->dev);
+	pci_dev_put(dev);
 	return 0;
 }
 
Signature · 1 line
2.18.4
author
#2 lore raw

PATCH v7 2/5 PCI/DPC: define a function to check and wait till port finish DPC handling

In reply to Ethan Zhao

Once root port DPC capability is enabled and triggered, at the beginning of DPC is triggered, the DPC status bits are set by hardware and then sends DPC/DLLSC/PDC interrupts to OS DPC and pciehp drivers, it will take the port and software DPC interrupt handler 10ms to 50ms (test data on ICS(Ice Lake SP platform, see
https://en.wikichip.org/wiki/intel/microarchitectures/ice_lake_(server)
& stable 5.9-rc6) to complete the DPC containment procedure till the DPC status is cleared at the end of the DPC interrupt handler.

We use this function to check if the root port is in DPC handling status and wait till the hardware and software completed the procedure.

---
changes:
v2:align ICS code name to public doc.
v3: no change.
v4: response to Christoph's (Christoph Hellwig <[email protected]>)
tip, move pci_wait_port_outdpc() to DPC driver and its declaration to pci.h.
v5: fix building issue reported by [email protected] with some config.
v6: move from [1/5] to [2/5].
v7: no change.

 drivers/pci/pci.h      |  2 ++
 drivers/pci/pcie/dpc.c | 27 +++++++++++++++++++++++++++
 2 files changed, 29 insertions(+)
drivers/pci/pci.h +2 -0
index fa12f7cbc1a0..455b32187abd 100644
@@ -455,10 +455,12 @@ void pci_restore_dpc_state(struct pci_dev *dev);
 void pci_dpc_init(struct pci_dev *pdev);
 void dpc_process_error(struct pci_dev *pdev);
 pci_ers_result_t dpc_reset_link(struct pci_dev *pdev);
+bool pci_wait_port_outdpc(struct pci_dev *pdev);
 #else
 static inline void pci_save_dpc_state(struct pci_dev *dev) {}
 static inline void pci_restore_dpc_state(struct pci_dev *dev) {}
 static inline void pci_dpc_init(struct pci_dev *pdev) {}
+static inline bool pci_wait_port_outdpc(struct pci_dev *pdev) { return false; }
 #endif
 
 #ifdef CONFIG_PCI_ATS
drivers/pci/pcie/dpc.c +27 -0
index daa9a4153776..2e0e091ce923 100644
@@ -71,6 +71,33 @@ void pci_restore_dpc_state(struct pci_dev *dev)
 	pci_write_config_word(dev, dev->dpc_cap + PCI_EXP_DPC_CTL, *cap);
 }
 
+bool pci_wait_port_outdpc(struct pci_dev *pdev)
+{
+	u16 cap = pdev->dpc_cap, status;
+	u16 loop = 0;
+
+	if (!cap) {
+		pci_WARN_ONCE(pdev, !cap, "No DPC capability initiated\n");
+		return false;
+	}
+	pci_read_config_word(pdev, cap + PCI_EXP_DPC_STATUS, &status);
+	pci_dbg(pdev, "DPC status %x, cap %x\n", status, cap);
+
+	while (status & PCI_EXP_DPC_STATUS_TRIGGER && loop < 100) {
+		msleep(10);
+		loop++;
+		pci_read_config_word(pdev, cap + PCI_EXP_DPC_STATUS, &status);
+	}
+
+	if (!(status & PCI_EXP_DPC_STATUS_TRIGGER)) {
+		pci_dbg(pdev, "Out of DPC %x, cost %d ms\n", status, loop*10);
+		return true;
+	}
+
+	pci_dbg(pdev, "Timeout to wait port out of DPC status\n");
+	return false;
+}
+
 static int dpc_wait_rp_inactive(struct pci_dev *pdev)
 {
 	unsigned long timeout = jiffies + HZ;
Signature · 1 line
2.18.4
author
#3 lore raw

PATCH v7 3/5 PCI: pciehp: check and wait port status out of DPC before handling DLLSC and PDC

In reply to Ethan Zhao

When root port has DPC capability and it is enabled, then triggered by errors, DPC DLLSC and PDC etc interrupts will be sent to DPC driver, pciehp drivers almost at the same time.
That will cause following messed and confused errors handling/recovery/removal /plugin procedure.

1. Port and device are in error recovery reseting initiated by DPC hardware, pciehp driver treats them as device is doing hot-remove or hot-plugin the same time.

2. While DPC handler calling device driver->err_handler callback( error_detected/resume etc), but the slot may be powered off by

pciehp
-> remove_board()

      -> pciehp_power_off_slot().

3. While DPC handler -> pci_do_recovery is doing different action to detect error and recover based on device->error_state, pciehp driver could change it on the fly by:

pciehp_unconfigure_device()
->pci_walk_bus()

     -> pci_dev_set_disconnected()

4. While DPC handler is calling device driver err_handler callback to detect error and recover, pciehp driver could is doing device unbind and release its driver.

...

While NON-FATAL/FATAL errors happen while hotplug is(is not)doing, result is not determinate.

So we need some kind of synchronization between pciehp DLLSC/PDC handling and DPC driver error recover handling. we need a determinate result of DPC error containment, link is recovered, link isn't recovered, device is still there, device is removed, then do pciehp hot-remove and hot-plugin procudure, don't mix them together.

Per our test on ICS platform, DPC error containment and software handler will take 10ms up to 50ms till clean the DPC triggered status. it is quick enough for pciehp compared with its 1000ms waiting to ignore DLLSC/PDC after doing power off.

With this patch, the handling flow of DPC containment and hotplug is partly ordered and serialized, let hardware DPC do the controller reset etc recovery action first, then DPC driver handling the call-back from device drivers, clear the DPC status, at the end, pciehp handle the DLLSC and PDC etc.

After tens of PCIe Gen4 NVMe SSD brute force hot-remove and hot-plugin with any time internval between the two actions, also stressed with the DPC injection test. system recovered to normal working state from NON-FATAL/FATAL errors as expected. hotplug works well without any random undeterminate errors or malfunction.

Brute DPC error injection script:

for i in {0..100}
do

        setpci -s 64:02.0 0x196.w=000a
        setpci -s 65:00.0 0x04.w=0544
        mount /dev/nvme0n1p1 /root/nvme
        sleep 1

done

---
Changes:
v2: revise doc according to Andy's suggestion.
v3: no change.
v4: no change.
v5: no change.
v6: moved to [3/5] from [2/5] and re-wrote description.
v7: no change.

 drivers/pci/hotplug/pciehp_hpc.c | 4 +++-
 1 file changed, 3 insertions(+), 1 deletion(-)
drivers/pci/hotplug/pciehp_hpc.c +3 -1
index 53433b37e181..6f271160f18d 100644
@@ -710,8 +710,10 @@ static irqreturn_t pciehp_ist(int irq, void *dev_id)
 	down_read(&ctrl->reset_lock);
 	if (events & DISABLE_SLOT)
 		pciehp_handle_disable_request(ctrl);
-	else if (events & (PCI_EXP_SLTSTA_PDC | PCI_EXP_SLTSTA_DLLSC))
+	else if (events & (PCI_EXP_SLTSTA_PDC | PCI_EXP_SLTSTA_DLLSC)) {
+		pci_wait_port_outdpc(pdev);
 		pciehp_handle_presence_or_link_change(ctrl, events);
+	}
 	up_read(&ctrl->reset_lock);
 
 	ret = IRQ_HANDLED;
Signature · 1 line
2.18.4
#4 lore raw

Re: [PATCH v7 3/5] PCI: pciehp: check and wait port status out of DPC before handling DLLSC and PDC

Quoted 3 lines · Ethan Zhao

When root port has DPC capability and it is enabled, then triggered by errors, DPC DLLSC and PDC etc interrupts will be sent to DPC driver, pciehp drivers almost at the same time.

Do the DLLSC and PDC events occur as a result of handling the error or do they occur independently?

If the latter, I don't see how we can tell whether the card in the slot is still the same.

If the former, holding the hotplug slot's reset_lock and doing something along the lines of pciehp_reset_slot() (or calling it directly) might solve the race.

Thanks,

Lukas

#5 lore raw

Re: [PATCH v7 3/5] PCI: pciehp: check and wait port status out of DPC before handling DLLSC and PDC

Lukas,

Quoted 8 lines · Lukas Wunner
Quoted 3 lines · Ethan Zhao

When root port has DPC capability and it is enabled, then triggered by errors, DPC DLLSC and PDC etc interrupts will be sent to DPC driver, pciehp drivers almost at the same time.

Do the DLLSC and PDC events occur as a result of handling the error or do they occur independently?

They could happen independently if links were recovered then the card was removed.
They could happen as a result of handling the errors the same time.

So don't assume DLLSC and PDC all occur at the same time.

Quoted 3 lines

If the latter, I don't see how we can tell whether the card in the slot is still the same.

If PDC happens, the card in the slot might not be the same. so hot-removal /hot -plugin handling follows the PDC event.

Quoted 4 lines

If the former, holding the hotplug slot's reset_lock and doing something along the lines of pciehp_reset_slot() (or calling it directly) might solve the race.

DPC reset is done by hardware, only AER calls pciehp_reset_slot() as recovery handling initiated by software.

Thanks,
Ethan

Quoted 4 lines

Thanks,

Lukas

author
#6 lore raw

PATCH v7 4/5 PCI: only return true when dev io state is really changed

In reply to Ethan Zhao

When uncorrectable error happens, AER driver and DPC driver interrupt handlers likely call

pcie_do_recovery()
->pci_walk_bus()

     ->report_frozen_detected()

with pci_channel_io_frozen the same time.
If pci_dev_set_io_state() return true even if the original state is
pci_channel_io_frozen, that will cause AER or DPC handler re-enter the error detecting and recovery procedure one after another.
The result is the recovery flow mixed between AER and DPC.
So simplify the pci_dev_set_io_state() function to only return true when dev->error_state is really changed.

---
Changnes:
v2: revise description and code according to suggestion from Andy.
v3: change code to simpler.
v4: no change.
v5: no change.
v6: no change.
v7: changed based on Bjorn's code and truth table.

 drivers/pci/pci.h | 53 ++++++++++++++++++-----------------------------
 1 file changed, 20 insertions(+), 33 deletions(-)
drivers/pci/pci.h +20 -33
index 455b32187abd..47af1ff2a286 100644
@@ -354,44 +354,31 @@ struct pci_sriov {
  *
  * Must be called with device_lock held.
  *
- * Returns true if state has been changed to the requested state.
+ * Returns true if state has been really changed to the requested state.
  */
 static inline bool pci_dev_set_io_state(struct pci_dev *dev,
 					pci_channel_state_t new)
 {
-	bool changed = false;
-
 	device_lock_assert(&dev->dev);
-	switch (new) {
-	case pci_channel_io_perm_failure:
-		switch (dev->error_state) {
-		case pci_channel_io_frozen:
-		case pci_channel_io_normal:
-		case pci_channel_io_perm_failure:
-			changed = true;
-			break;
-		}
-		break;
-	case pci_channel_io_frozen:
-		switch (dev->error_state) {
-		case pci_channel_io_frozen:
-		case pci_channel_io_normal:
-			changed = true;
-			break;
-		}
-		break;
-	case pci_channel_io_normal:
-		switch (dev->error_state) {
-		case pci_channel_io_frozen:
-		case pci_channel_io_normal:
-			changed = true;
-			break;
-		}
-		break;
-	}
-	if (changed)
-		dev->error_state = new;
-	return changed;
+
+/*
+ *			Truth table:
+ *			requested new state
+ *     current          ------------------------------------------
+ *     state            normal         frozen         perm_failure
+ *     ------------  +  -------------  -------------  ------------
+ *     normal        |  normal         frozen         perm_failure
+ *     frozen        |  normal         frozen         perm_failure
+ *     perm_failure  |  perm_failure*  perm_failure*  perm_failure
+ */
+
+	if (dev->error_state == pci_channel_io_perm_failure)
+		return false;
+	else if (dev->error_state == new)
+		return false;
+
+	dev->error_state = new;
+	return true;
 }
 
 static inline int pci_dev_set_disconnected(struct pci_dev *dev, void *unused)
Signature · 1 line
2.18.4
#7 lore raw

Re: [PATCH v7 4/5] PCI: only return true when dev io state is really changed

Quoted 88 lines · Ethan Zhao

When uncorrectable error happens, AER driver and DPC driver interrupt handlers likely call

   pcie_do_recovery()
   ->pci_walk_bus()
     ->report_frozen_detected()

with pci_channel_io_frozen the same time.
If pci_dev_set_io_state() return true even if the original state is
pci_channel_io_frozen, that will cause AER or DPC handler re-enter the error detecting and recovery procedure one after another.
The result is the recovery flow mixed between AER and DPC.
So simplify the pci_dev_set_io_state() function to only return true when dev->error_state is really changed.

---
Changnes:
v2: revise description and code according to suggestion from Andy.
v3: change code to simpler.
v4: no change.
v5: no change.
v6: no change.
v7: changed based on Bjorn's code and truth table.

 drivers/pci/pci.h | 53 ++++++++++++++++++-----------------------------
 1 file changed, 20 insertions(+), 33 deletions(-)
drivers/pci/pci.h +10 -33
index 455b32187abd..47af1ff2a286 100644
@@ -354,44 +354,31 @@ struct pci_sriov {
  *
  * Must be called with device_lock held.
  *
- * Returns true if state has been changed to the requested state.
+ * Returns true if state has been really changed to the requested state.
  */
 static inline bool pci_dev_set_io_state(struct pci_dev *dev,
 					pci_channel_state_t new)
 {
-	bool changed = false;
-
 	device_lock_assert(&dev->dev);
-	switch (new) {
-	case pci_channel_io_perm_failure:
-		switch (dev->error_state) {
-		case pci_channel_io_frozen:
-		case pci_channel_io_normal:
-		case pci_channel_io_perm_failure:
-			changed = true;
-			break;
-		}
-		break;
-	case pci_channel_io_frozen:
-		switch (dev->error_state) {
-		case pci_channel_io_frozen:
-		case pci_channel_io_normal:
-			changed = true;
-			break;
-		}
-		break;
-	case pci_channel_io_normal:
-		switch (dev->error_state) {
-		case pci_channel_io_frozen:
-		case pci_channel_io_normal:
-			changed = true;
-			break;
-		}
-		break;
-	}
-	if (changed)
-		dev->error_state = new;
-	return changed;
+
+/*
+ *			Truth table:
+ *			requested new state
+ *     current          ------------------------------------------
+ *     state            normal         frozen         perm_failure
+ *     ------------  +  -------------  -------------  ------------
+ *     normal        |  normal         frozen         perm_failure
+ *     frozen        |  normal         frozen         perm_failure

drivers/pci/pci.h struct pci_sriov { · line 381

Quoted 10 lines
drivers/pci/pci.h +10 -0
+ *     perm_failure  |  perm_failure*  perm_failure*  perm_failure
+ */
+
+	if (dev->error_state == pci_channel_io_perm_failure)
+		return false;
+	else if (dev->error_state == new)
+		return false;
+
+	dev->error_state = new;
+	return true;

No, you missed the point. I want

1) One patch that converts the "switch" to the shorter "if" statements. This one will be big and ugly, but should not change the functionality at all, and it should be pretty easy to verify that since there aren't very many states involved.

Since this one is pure code simplification, the commit log won't say anything at all about AER or DPC or their requirements because it's not changing any behavior.

2) A separate patch that's tiny and makes whatever functional change you need.

Quoted 6 lines
drivers/pci/pci.h +0 -0
 }
 
 static inline int pci_dev_set_disconnected(struct pci_dev *dev, void *unused)
Signature · 1 line
2.18.4
#8 lore raw

Re: [PATCH v7 4/5] PCI: only return true when dev io state is really changed

Bjorn,

Quoted 108 lines · Bjorn Helgaas
Quoted 88 lines · Ethan Zhao

When uncorrectable error happens, AER driver and DPC driver interrupt handlers likely call

   pcie_do_recovery()
   ->pci_walk_bus()
     ->report_frozen_detected()

with pci_channel_io_frozen the same time.
If pci_dev_set_io_state() return true even if the original state is
pci_channel_io_frozen, that will cause AER or DPC handler re-enter the error detecting and recovery procedure one after another.
The result is the recovery flow mixed between AER and DPC.
So simplify the pci_dev_set_io_state() function to only return true when dev->error_state is really changed.

---
Changnes:
v2: revise description and code according to suggestion from Andy.
v3: change code to simpler.
v4: no change.
v5: no change.
v6: no change.
v7: changed based on Bjorn's code and truth table.

 drivers/pci/pci.h | 53 ++++++++++++++++++-----------------------------
 1 file changed, 20 insertions(+), 33 deletions(-)
drivers/pci/pci.h +10 -33
index 455b32187abd..47af1ff2a286 100644
@@ -354,44 +354,31 @@ struct pci_sriov {
  *
  * Must be called with device_lock held.
  *
- * Returns true if state has been changed to the requested state.
+ * Returns true if state has been really changed to the requested state.
  */
 static inline bool pci_dev_set_io_state(struct pci_dev *dev,
                                      pci_channel_state_t new)
 {
-     bool changed = false;
-
      device_lock_assert(&dev->dev);
-     switch (new) {
-     case pci_channel_io_perm_failure:
-             switch (dev->error_state) {
-             case pci_channel_io_frozen:
-             case pci_channel_io_normal:
-             case pci_channel_io_perm_failure:
-                     changed = true;
-                     break;
-             }
-             break;
-     case pci_channel_io_frozen:
-             switch (dev->error_state) {
-             case pci_channel_io_frozen:
-             case pci_channel_io_normal:
-                     changed = true;
-                     break;
-             }
-             break;
-     case pci_channel_io_normal:
-             switch (dev->error_state) {
-             case pci_channel_io_frozen:
-             case pci_channel_io_normal:
-                     changed = true;
-                     break;
-             }
-             break;
-     }
-     if (changed)
-             dev->error_state = new;
-     return changed;
+
+/*
+ *                   Truth table:
+ *                   requested new state
+ *     current          ------------------------------------------
+ *     state            normal         frozen         perm_failure
+ *     ------------  +  -------------  -------------  ------------
+ *     normal        |  normal         frozen         perm_failure
+ *     frozen        |  normal         frozen         perm_failure

drivers/pci/pci.h struct pci_sriov { · line 381

Quoted 10 lines
drivers/pci/pci.h +10 -0
+ *     perm_failure  |  perm_failure*  perm_failure*  perm_failure
+ */
+
+     if (dev->error_state == pci_channel_io_perm_failure)
+             return false;
+     else if (dev->error_state == new)
+             return false;
+
+     dev->error_state = new;
+     return true;

No, you missed the point. I want

1) One patch that converts the "switch" to the shorter "if" statements. This one will be big and ugly, but should not change the functionality at all, and it should be pretty easy to verify that since there aren't very many states involved.

Quoted 6 lines

Since this one is pure code simplification, the commit log won't say anything at all about AER or DPC or their requirements because it's not changing any behavior.

2) A separate patch that's tiny and makes whatever functional change you need.

Make sense, clear, this time.

Thanks,
Ethan

Quoted 7 lines
Quoted 7 lines
drivers/pci/pci.h +0 -1
 }
 
 static inline int pci_dev_set_disconnected(struct pci_dev *dev, void *unused)
--

2.18.4

author
#9 lore raw

PATCH v7 5/5 PCI/ERR: don't mix io state not changed and no driver together

In reply to Ethan Zhao

When we see 'can't recover (no error_detected callback)' on console, Maybe the reason is io state is not changed by calling pci_dev_set_io_state(), that is confused. fix it.

---
Chagnes:
v2: no change.
v3: no change.
v4: no change.
v5: no change.
v6: no change.
v7: change debug output information.

 drivers/pci/pcie/err.c | 6 ++++--
 1 file changed, 4 insertions(+), 2 deletions(-)
drivers/pci/pcie/err.c +4 -2
index e35c4480c86b..2ca2723f3b34 100644
@@ -55,8 +55,10 @@ static int report_error_detected(struct pci_dev *dev,
 	if (!pci_dev_get(dev))
 		return 0;
 	device_lock(&dev->dev);
-	if (!pci_dev_set_io_state(dev, state) ||
-		!dev->driver ||
+	if (!pci_dev_set_io_state(dev, state)) {
+		pci_dbg(dev, "Device was in that state or not allowed setting.\n");
+		vote = PCI_ERS_RESULT_NONE;
+	} else if (!dev->driver ||
 		!dev->driver->err_handler ||
 		!dev->driver->err_handler->error_detected) {
 		/*
Signature · 1 line
2.18.4
#10 lore raw
In reply to Ethan Zhao

Hi Ethan

Quoted 8 lines · Ethan Zhao

Hi,folks,

This simple patch set fixed some serious security issues found when DPC error injection and NVMe SSD hotplug brute force test were doing -- race condition between DPC handler and pciehp, AER interrupt handlers, caused system hang and system with DPC feature couldn't recover to normal working state as expected (NVMe instance lost, mount operation hang, race PCIe access caused uncorrectable errors reported alternatively etc).

I think maybe picking from other commit messages to make this description in cover letter bit clear. The fundamental premise is that when due to error conditions when events are processed by both DPC handler and hotplug handling of DLLSC both operating on the same device object ends up with crashes.

Quoted 67 lines

With this patch set applied, stable 5.9-rc6 on ICS (Ice Lake SP platform, see
https://en.wikichip.org/wiki/intel/microarchitectures/ice_lake_(server))

could pass the PCIe Gen4 NVMe SSD brute force hotplug test with any time interval between hot-remove and plug-in operation tens of times without any errors occur and system works normal.

With this patch set applied, system with DPC feature could recover from NON-FATAL and FATAL errors injection test and works as expected.

System works smoothly when errors happen while hotplug is doing, no uncorrectable errors found.

Brute DPC error injection script:

for i in {0..100}
do

        setpci -s 64:02.0 0x196.w=000a
        setpci -s 65:00.0 0x04.w=0544
        mount /dev/nvme0n1p1 /root/nvme
        sleep 1

done

Other details see every commits description part.

This patch set could be applied to stable 5.9-rc6/rc7 directly.

Help to review and test.

v2: changed according to review by Andy Shevchenko.
v3: changed patch 4/5 to simpler coding.
v4: move function pci_wait_port_outdpc() to DPC driver and its
declaration to pci.h. (tip from Christoph Hellwig <[email protected]>).
v5: fix building issue reported by [email protected] with some config.
v6: move patch[3/5] as the first patch according to Lukas's suggestion.
and rewrite the comment part of patch[3/5].
v7: change the patch[4/5], based on Bjorn's code and truth table.
change the patch[5/5] about the debug output information.

Thanks,
Ethan

Ethan Zhao (5):
PCI/ERR: get device before call device driver to avoid NULL pointer
dereference
PCI/DPC: define a function to check and wait till port finish DPC
handling
PCI: pciehp: check and wait port status out of DPC before handling
DLLSC and PDC
PCI: only return true when dev io state is really changed PCI/ERR: don't mix io state not changed and no driver together

 drivers/pci/hotplug/pciehp_hpc.c |  4 ++-
 drivers/pci/pci.h                | 55 +++++++++++++-------------------
 drivers/pci/pcie/dpc.c           | 27 ++++++++++++++++
 drivers/pci/pcie/err.c           | 18 +++++++++--
 4 files changed, 68 insertions(+), 36 deletions(-)

base-commit: a1b8638ba1320e6684aa98233c15255eb803fac7

Signature · 1 line
2.18.4
#11 lore raw

Raj,

Quoted 13 lines · Ashok

Hi Ethan

Quoted 8 lines · Ethan Zhao

Hi,folks,

This simple patch set fixed some serious security issues found when DPC error injection and NVMe SSD hotplug brute force test were doing -- race condition between DPC handler and pciehp, AER interrupt handlers, caused system hang and system with DPC feature couldn't recover to normal working state as expected (NVMe instance lost, mount operation hang, race PCIe access caused uncorrectable errors reported alternatively etc).

Quoted 4 lines

I think maybe picking from other commit messages to make this description in cover letter bit clear. The fundamental premise is that when due to error conditions when events are processed by both DPC handler and hotplug handling of DLLSC both operating on the same device object ends up with crashes.

Yep, that's right.

Thanks,
Ethan

Quoted 69 lines
Quoted 69 lines

With this patch set applied, stable 5.9-rc6 on ICS (Ice Lake SP platform, see
https://en.wikichip.org/wiki/intel/microarchitectures/ice_lake_(server))

could pass the PCIe Gen4 NVMe SSD brute force hotplug test with any time interval between hot-remove and plug-in operation tens of times without any errors occur and system works normal.

With this patch set applied, system with DPC feature could recover from NON-FATAL and FATAL errors injection test and works as expected.

System works smoothly when errors happen while hotplug is doing, no uncorrectable errors found.

Brute DPC error injection script:

for i in {0..100}
do

        setpci -s 64:02.0 0x196.w=000a
        setpci -s 65:00.0 0x04.w=0544
        mount /dev/nvme0n1p1 /root/nvme
        sleep 1

done

Other details see every commits description part.

This patch set could be applied to stable 5.9-rc6/rc7 directly.

Help to review and test.

v2: changed according to review by Andy Shevchenko.
v3: changed patch 4/5 to simpler coding.
v4: move function pci_wait_port_outdpc() to DPC driver and its
declaration to pci.h. (tip from Christoph Hellwig <[email protected]>).
v5: fix building issue reported by [email protected] with some config.
v6: move patch[3/5] as the first patch according to Lukas's suggestion.
and rewrite the comment part of patch[3/5].
v7: change the patch[4/5], based on Bjorn's code and truth table.
change the patch[5/5] about the debug output information.

Thanks,
Ethan

Ethan Zhao (5):
PCI/ERR: get device before call device driver to avoid NULL pointer
dereference
PCI/DPC: define a function to check and wait till port finish DPC
handling
PCI: pciehp: check and wait port status out of DPC before handling
DLLSC and PDC
PCI: only return true when dev io state is really changed PCI/ERR: don't mix io state not changed and no driver together

 drivers/pci/hotplug/pciehp_hpc.c |  4 ++-
 drivers/pci/pci.h                | 55 +++++++++++++-------------------
 drivers/pci/pcie/dpc.c           | 27 ++++++++++++++++
 drivers/pci/pcie/err.c           | 18 +++++++++--
 4 files changed, 68 insertions(+), 36 deletions(-)

base-commit: a1b8638ba1320e6684aa98233c15255eb803fac7 --
2.18.4